Still, some executives at pharma companies worry about AI’s potential to generate misleading data, as the Economist notes. Such fears are not entirely unfounded. While most headlines continue to tout AI’s promise in the field, there was a notable failure in 2023: an AI-aided drug failed a pair of phase 3 trials.
This article provides an overview of AI’s increasing traction in drug discovery and development in 2023 from January to early August. We’ll update it as the year progresses.
January 2023: Absci deploys zero-shot generative AI in antibody design
Type of AI: Early this year, the generative AI drug discovery firm Absci revealed that it had designed and validated de novo therapeutic antibodies with zero-shot generative AI. Zero-shot learning refers to a model that can make accurate decisions — here, in the domain of antibody design — without any training data.
Why it matters: The company said the breakthrough could slash the time in half to get new drug candidates to the clinic. Absci also claims the technique could improve the odds of such therapeutic antibodies of making it through clinical trials. It posted results on the preprint site BioRxiv.
March 2023: MIT reveals DiffDock, which could support faster, safer drug development
Type of AI: A significant development in AI in drug discovery 2023 came with the revealing of MIT’s DiffDock, which could support faster, safer drug development. DiffDock is based on diffusion generative models, specifically used in the context of molecular docking.
Why it matters: DiffDock offers a new method for predicting molecular docking, or how a small molecule, or ligand, will bind to a protein. Specifically, DiffDock maps the degrees of freedom involved in docking. In tests, DiffDock had a 38% success rate, which is higher than traditional docking prediction methods. Those have a roughly 23% success rate. Deep learning methods score even lower, with a 20% success rate. The research was also featured in a preprint.
March 2023: Nvidia launches BioNeMo Cloud to boost drug discovery with generative AI
Type of AI: Nvidia’s BioNeMo Cloud is a cloud service for generative AI–based drug discovery.Why it matters: Nvidia notes that its cloud APIs enable researchers to customize and deploy domain-specific, generative and predictive biomolecular AI models at scale. Researchers can fine-tune models with proprietary data, executing inference via the web or APIs.
AstraZeneca and several startups, including Insilico Medicine and Evozyne, are BioNeMo customers. AstraZeneca is partnering with Nvidia to use the platform in tandem with supercomputers. Earlier this year, Evozyne announced that it had used the service to develop ‘“supernatural” proteins — that is, therapeutic proteins that are potentially more effective than naturally occurring versions for phenylketonuria, a rare metabolic disorder.
June 2023: Insilico Medicine’s AI drug enters phase 2 study
Type of AI: Insilico used generative AI to discover INS018_055, a small molecule which recently entered a phase 2 study as a potential therapy for idiopathic pulmonary fibrosis. The company uses NVIDIA Tensor Core GPUs in its generative AI drug design engine, Chemistry42, to generate novel molecular structures. Overall, the company’s Pharma.AI platform makes use of multiple AI models. Here, the company used generative AI to identify a molecule that a drug compound could target, generate novel drug candidates and estimate how well these candidates would bind with the target. The company has also succeeded in using generative AI to predict clinical trial outcomes.
Why it matters: INS018_055 is the first AI-discovered drug to reach phase 2 trials. The company estimates that, with traditional methods, the drug candidate would have cost more than $400 million and taken up to six years to develop. But with generative AI, it reached those objectives at one-tenth of the cost and in one-third of the time, Insilico estimates. In an interview with Drug Discovery & Development, the company’s CEO noted that the company was able to go from novel target discovery to phase 1 in fewer than 30 months.
June 2023: Sanofi reveals plan to put AI at the center of its operations
Type of AI: In June, Sanofi announced its plans for a large-scale implementation of an AI-driven app known as Plai that it developed with Aily Labs. The app makes use of internal company data to offer real-time insights and personalized scenarios to support data-driven decision-making or “nudges” as CEO Paul Hudson described them in an interview with the Economist.
Why it matters: The use of AI has enabled Sanofi to accelerate research processes, improve clinical trial design, optimize manufacturing and drive more efficiency in its supply chain. The company has also forged partnerships with several companies to speed the development of therapies for cancer and immune system-related diseases.
While Sanofi appears to be ambitious in its quest to deploy AI across its operations, it is not alone. AstraZeneca, Moderna and other Big Pharmas have comprehensive AI and data science initiatives.
July 2023: AI-aided drug ulotaront fails phase 3 studies
Type of AI: Sumitomo Pharma and PsychoGenics developed the schizophrenia drug candidate ulotaront using the latter’s SmartCube platform, which is an automated testing platform. SmartCube works by presenting a series of challenges to a mouse through custom hardware. The platform then uses computer vision to process and analyze large temporal and vectoral datasets. SmartCube is essentially “a robot where the mouse is placed inside,” explained Daniela Brunner, chief innovation officer of PsychoGenics, in an interview earlier this year.
Why it matters: While AI can help reduce drug development risk, the technology doesn’t completely eliminate it. Despite promising phase 2 results, ulotaront failed to outperform placebo in DIAMOND 1 and DIAMOND 2, two large phase 3 studies of acutely psychotic adults with schizophrenia. Sumitomo and Otsuka, ulotaront’s developers, have not yet signaled that they are giving up on the trace amine-associated receptor (TAAR) agonist ulotaront. The companies state that high placebo response is common in schizophrenia trials, and are mulling next steps with the FDA.
July 2023: Nvidia invests $50 million in Recursion to boost AI-driven drug discovery
Type of AI: In July 2023, Nvidia revealed it was making a $50 million investment in the biotech Recursion to support the training of the firm’s AI models for drug discovery. Recursion is using a variety of advanced technologies, including convolutional neural networks for sifting through high-dimensional biological and chemical data. It also deploys data processing tools, biological and chemical activity assessment software to shed light on how compounds act and AI-powered decision support tools. Additionally, it makes use of a highly scalable, automated laboratory robotics workflow.
Why it matters: Nvidia plans to license Recursion’s models to biotech firms through BioNeMo, its generative AI cloud service for drug discovery launched earlier in the year that is described earlier. Nvidia now has a 4% stake in Recursion, further underscoring the growing role of Big Tech firms in the life sciences. Recursion also plans to use Nvidia’s software to support its internal drug development efforts as well as the pipelines of partners like Bayer and Roche.
July 2023: AI startup Causaly raises $60 million to expedite drug development
Type of AI: Causaly is using a combination of knowledge graphs and generative AI in its platform. In addition, the company is using generative AI to automate traditionally labor-intensive processes of biomedical research and discovery.
Why it matters: London-based startup Causaly has developed an AI platform to accelerate the development and testing of drugs. In July, the company revealed that it had raised $60 million in a Series B funding. ICONIQ Growth led the funding round.
Big Pharma heavyweights like Gilead, Novo Nordisk and Regeneron are Causaly customers. So are the FDA and National Institute of Environmental Health Sciences.
August 2023: Pharos iBio develops anticancer drug with AI platform
Type of AI: Pharos iBio’s AI drug discovery platform, Chemiverse, uses a variety of AI techniques to identify and develop targeted anticancer drugs. The AI platform uses machine learning and deep learning algorithms to find patterns in large drug discovery datasets. The platform can simulate and analyze protein-ligand interactions, which are pivotal for understanding a drug’s potential safety and efficacy. Additionally, the Chemiverse platform uses molecular interaction energy calculations to shed light on the binding affinity between a prospective drug and its target protein.
Why it matters: With Chemiverse, Pharos iBio successfully identified and developed PHI-101, a targeted anticancer drug that combats the FLT3 gene mutation found in roughly 30% to 35% of patients with acute myeloid leukemia (AML).
Already, Korea’s Ministry of Food and Drug Safety has granted the approval to the drug on an individual basis, allowing a healthcare provider at the Catholic University of Korea-Seoul St. Mary’s Hospital to administer it to an eligible patient. Researchers are also testing the drug candidate’s potential in platinum-resistant recurrent ovarian cancer.
August 2023: Generative AI tool boasts 79% accuracy in predicting clinical trial outcomes
Type of AI: In August 2023, Insilico Medicine announced a significant breakthrough in predicting clinical trial outcomes with its generative AI tool, InClinico. The tool uses transformer-based AI models and multimodal data sources, including text, omics, trial design and drug properties. The company trained the system on more than 55,600 phase 2 clinical trials.
Why it matters: Developed over seven years, InClinico’s accuracy rate in forecasting the outcomes of real-world phase 2 and 3 trials could lay the groundwork for improving drug development efficiency. Results of the study were published in the peer-reviewed Clinical Pharmacology and Therapeutics. Insilico also noted that the platform could serve as a valuable tool for investors, supporting a 35% 9-month return on investment in a virtual trading portfolio.
August 2023: UCF researchers reveal AI-assisted drug screening technology
Type of AI: The BindingSite-AugmentedDTA model from the University of Central Florida is a deep learning (DL)-based framework. It aims to enhance drug–target affinity (DTA) predictions by optimizing the search for potential-binding sites of proteins. The model is based on an augmented-DTA in-silico prediction module that uses a graph convolutional neural network (GCNN)-based model known as AttentionSiteDTI. The “DTI” in the name refers to drug-target interactions. The model serves as a detector to identify the most likely binding sites of the target protein.
Why it matters: The researchers, who published the research in Briefing in Bioinformatics, note that the model is highly generalizable and adaptable, given its potential for integration with any deep learning-based regression model. With the potential to make binding affinity prediction more efficient and accurate, the framework can bolster the accuracy of drug target affinity predictions. The researchers have already validated the prediction power of the model through in vitro experiments. Next, the research team will continue validating the model’s prediction capabilities with a variety of drugs and proteins.
Recursion uses AI to bridge the protein and chemical universe, predicting targets for 36 billion compounds
Type of AI: A month after receiving an injection of funding from Nvidia (see above), Recursion Pharmaceuticals announced a breakthrough related to using machine learning to predict the binding sites of proteins. Specifically, Recursion is using Cyclica’s MatchMaker technology, NVIDIA DGX Cloud supercomputing and DeepMind’s AlphaFold2 database to screen Enamine REAL Space, a searchable chemical library, to predict the protein targets for 36 billion chemical compounds. This approach unites high-performance computing with expansive chemical databases, where machine learning determines binding compatibility between a small molecule and a protein binding pocket. Recursion says this is more efficient than traditional docking and physics-based interaction simulations.
Why it matters: Recursion noted that the combination of technologies enabled research in a single week that would have otherwise taken 100,000 years using traditional methods. The company was able to predict the protein targets for approximately 36 billion chemical compounds, a major feat. To achieve that milestone, Recursion digitally evaluated more than 2.8 quadrillion small molecule-target pairs, representing a massive computational process. The process paired Recursion’s multimodal dataset to yield a comprehensive map of how these chemicals could potentially interact with bodily proteins. These protein-ligand interaction data can inform wet-lab experiments and accelerate medicinal chemistry cycles.
Filed Under: clinical trials, Data science, Drug Discovery, Industry 4.0, machine learning and AI